Model availability tracker
Gemini 4 Carbon release date: status, price and availability
Gemini 4 Carbon was announced October 28, 2026 and is not generally available yet — cyber defenders get it first via Google's Fairwind Program. The full status timeline, introductory pricing, the vendor-published benchmark set, and what the numbers mean for typed decision routing. Updated November 1, 2026.
TL;DR
Gemini 4 Carbon was announced on October 28, 2026. You probably cannot use it yet: access is rolling out to a subset of cyber defenders through Google's Fairwind Program and to trusted partners, while Google goes through the U.S. government's voluntary pre-release testing process. There is no public release date — Google's entire commitment is "as soon as possible". Pricing is announced: $2 per million input tokens and $10 per million output tokens introductory (same as Argon's intro), rising to $4 / $20 after. Vendor-published benchmarks put it #1 on the Artificial Analysis Intelligence Index v5.0 (59) with strong security-task numbers — treat those as claims to verify on your own tasks, not as verdicts.
Sources: Google DeepMind announcement (blog.google, 2026-10-28) and Reuters launch coverage, both retrieved 2026-11-01. Every number carries its as-of date because every row of this page moves.
The direct answer
Three things most searches actually want, each with its as-of date.
Status (as of 2026-11-01)
- Announced 2026-10-28 by Google DeepMind as its "next frontier model".
- Not generally available — Fairwind Program cyber defenders and trusted partners only.
- In the U.S. government's voluntary pre-release model testing process.
Price (announced at launch)
- Introductory: $2 / 1M input tokens, $10 / 1M output tokens — identical to Argon's intro.
- Cached input at 95% off the input price.
- After the introductory period: $4 / $20 per 1M — no end date published.
What to watch
- A public GA date — the announcement only says developers get it "as soon as possible".
- The $4 / $20 flip and whether it lands before or after GA.
- Argon's priority tier going to $36 / 1M output on 2026-12-01 — a preview of how fast intro pricing moves.
Release timeline
From Argon to Carbon to whatever "as soon as possible" turns out to mean.
| Date | Event | Status |
|---|---|---|
| 2026-09-30 | Gemini 4 Argon announced — the previous Gemini 4 flagship, at the same introductory $2 / $10 pricing. | Shipped |
| 2026-10-28 | Gemini 4 Carbon announced by Google DeepMind as "our next frontier model" — cybersecurity across the defense lifecycle (detection, validation, remediation) and long-horizon agentic workflows headlined. | Shipped |
| 2026-10-28 → | Access rolling out to a subset of cyber defenders via Google's Fairwind Program, then to additional trusted partners. Google is also in the U.S. government's voluntary pre-release model testing process. | Rolling out |
| TBA | General availability for developers, businesses and consumers: Google says "as soon as possible" — no date, no waitlist mechanics published as of 2026-11-01. This row is the reason the page exists; we update it when the date lands. | No date |
Statuses as of 2026-11-01. The "TBA" row is updated when Google publishes a date — that is the promise of this page.
Pricing: announced, and already on a clock
Introductory Gemini 4 pricing is a countdown, not a constant — Argon's priority tier is the proof.
| Item | Price | Note |
|---|---|---|
| Carbon — input (introductory) | $2 / 1M tokens | Same introductory input price as Argon. |
| Carbon — output (introductory) | $10 / 1M tokens | Cached input billed at 95% off the input price. |
| Carbon — after the introductory period | $4 / $20 per 1M | Announced at launch; Google has not published when the intro period ends. |
| Argon — priority tier output | $36 / 1M tokens | From 2026-12-01 — the concrete example of how fast introductory Gemini 4 pricing moves. |
All prices per million tokens as published on 2026-10-28. "Introductory period" has no published end date; Argon's $36 / 1M priority-tier output from 2026-12-01 shows what happens when it ends.
The benchmark set Google published
Security-heavy, vendor-selected, and useful — as a checklist of what to verify on your own tasks.
| Benchmark | What it measures | Published result |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | Composite capability index (third-party) | #1 (59) |
| SEC-bench Pro | Vulnerability finding | 88.2% |
| ExploitGym | Exploitation | 54.8% |
| CWE-bench (Collinear) | Patching | 71.0% |
| WebDev Arena | Full-stack web dev, Elo-ranked | #4 (1768) |
| Terminal-Bench 4.0 | Agentic terminal work — successful cheating | 0.0% (attempts 0.4%) |
| Gray Swan indirect prompt injection | Attack success rate (Argon: 0.7%) | 0.0% |
These are launch-day numbers from Google's announcement (2026-10-28): selected by the vendor, partly measured by a ratings service (Artificial Analysis), and never a substitute for running your own task-level experiments. That is not a dig — it is the same audit rule we apply to TypeSafe's homepage claims, which measured 3x and 12x against advertised 193x and 444x.
What this means for typed AI decisions
For routed, gated, typed workloads the release changes the flagship menu — not the arithmetic.
If your architecture sends every small judgment to a frontier model, Carbon's launch is a price event: the intro $2 / $10 matches Argon, and both step up later. But the architecture itself is the cost driver. A typed decision layer keeps the small calls off the frontier menu entirely — hosted Jev advertises $0.042 per million input tokens, so a 212-token decision input costs roughly $0.000009 there against about $0.00042 at Carbon's introductory input price: about 47x per call before output tokens, and the ratio only widens at the $4 / $20 list price.
The safety numbers are the more interesting signal for agentic pipelines. Gray Swan indirect prompt injection at 0.0% (Argon: 0.7%) and Terminal-Bench 4.0 successful cheating at 0.0% are exactly the failure modes a harness gates against — and if they hold up, a Carbon-backed agent needs fewer defensive gates around its tool calls, not fewer measurements. Verify on your own payloads first; launch benchmarks are claims until re-run on your tasks.
Practical posture while there is no GA date: keep the router pattern — a cheap typed classifier decides small-vs-flagship before each run — and treat Carbon as a new entry in the flagship column when it lands. Nothing about the pattern changes; only the flagship's price and capability row does. When you do evaluate it, benchmark it the way our benchmarks section does: fixed tasks, fixed thresholds, your data, your confidence gates.
// Per-call input cost at 212 tokens (the Jev audit-record example) carbonIntro = 212 / 1_000_000 * $2.00 // $0.000424 carbonList = 212 / 1_000_000 * $4.00 // $0.000848 hostedJev = 212 / 1_000_000 * $0.042 // $0.0000089 ratioVsIntro = carbonIntro / hostedJev // ≈ 47.6x ratioVsList = carbonList / hostedJev // ≈ 95x // Output tokens not included — and output is where frontier pricing bites.
Prices as published 2026-10-28 (Carbon intro and post-intro) and in Argon launch coverage (Jev hosted input price from TypeSafe's published pricing page). Math shown, not asserted — re-run it with your own token counts.
Update log
Page published: announcement 2026-10-28, Fairwind rollout, introductory $2 / $10 pricing (→ $4 / $20), vendor-published benchmark set, no GA date.
Newest first. If a row here contradicts your memory, trust the log and check the announcement — then tell us what we missed.
FAQ
When will Gemini 4 Carbon be released?
No public date exists as of November 1, 2026. Google's announcement (October 28, 2026) commits only to "as soon as possible" for developers, businesses and consumers, with no waitlist mechanics published. The rollout in progress is cyber-defender-first via the Fairwind Program.
Can I use Gemini 4 Carbon now?
Only if you are in the Fairwind Program's current subset of cyber defenders or one of Google's trusted partners. Google is also running the U.S. government's voluntary pre-release model testing process. Everyone else waits for general availability.
How much will Gemini 4 Carbon cost?
Announced at launch: $2 per million input tokens and $10 per million output tokens introductory — the same introductory pricing as Gemini 4 Argon — with cached input at 95% off. After the introductory period it rises to $4 / $20 per million. Google has not said when the introductory period ends.
What is the difference between Gemini 4 Carbon and Argon?
Carbon is positioned as the stronger successor, built on Argon (announced September 30, 2026) with the same safety stack and the same introductory pricing at launch. The headline differences are cybersecurity depth across detection, validation and remediation, long-horizon agentic work, and vendor-published numbers like #1 on the Artificial Analysis Intelligence Index v5.0 (59) and 0.0% on Gray Swan's indirect prompt injection benchmark (Argon: 0.7%).
Is Gemini 4 Carbon good for coding and agents?
The published numbers point that way: WebDev Arena #4 (1768 Elo), Terminal-Bench 4.0 with 0.0% successful cheating, and internal claims of 2x+ uplift over Argon on secure code generation. These are launch-day claims selected by the vendor — worth testing on your own repo and tool loop before re-architecting anything around them.
Should I route my AI decisions through Gemini 4 Carbon when it ships?
Route generation to whichever flagship wins your own evals — but keep small, high-frequency judgments (routing, gating, classification) on a typed decision layer. That arithmetic is independent of which flagship is newest: at introductory pricing Carbon's input still costs ~47x more per 212-token call than hosted Jev, before output tokens.